CN108764250A - A method of extracting essential image with convolutional neural networks - Google Patents
A method of extracting essential image with convolutional neural networks Download PDFInfo
- Publication number
- CN108764250A CN108764250A CN201810407424.8A CN201810407424A CN108764250A CN 108764250 A CN108764250 A CN 108764250A CN 201810407424 A CN201810407424 A CN 201810407424A CN 108764250 A CN108764250 A CN 108764250A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- branch
- neural networks
- proprietary
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention provides a kind of methods for extracting essential image with convolutional neural networks.First, a double-current convolutional network with parallel construction from image to image is built;Then, it is trained using specific training data set pair network, optimizes network parameter, to extract the multilayer feature with environment invariance, directly reconstruct essential image (reflectogram and illumination pattern).There is powerful ability in feature extraction as a result of the double-current convolutional neural networks based on deep learning the Theory Construction, reflectogram and illumination pattern can be directly isolated from original image.Simultaneously, the model is a kind of full convolutional network model from image to image, including Liang Ge branches flow to, it is respectively used to generate illumination pattern and reflectogram, and result of the convolution results of higher level after deconvolution operation is combined by the network structure, the reconstructed error that illumination pattern and reflectogram can be reduced to a certain extent improves the ability of network characterization reconstruct.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to a kind of side extracting essential image with convolutional neural networks
Method.
Background technology
The understanding of image and analysis are most basic one of the tasks of image procossing.Since image imaging process is by many factors
Such as joint effect of target object self character, shooting environmental and collecting device condition so that image processing process needs
Fully consider the interference of factors, such as the transformation of the discontinuity of shade, color, targeted attitude.These changing factors pair
Image processing algorithm brings larger challenge so that existing image analysis algorithm under complex environment performance by larger
It influences.Therefore, the robustness for image analysis algorithm how being improved under complex environment has become research hotspot in recent years.It is true
On, if it is possible to it is based on existing observed image, the substantive characteristics in image is analyzed, image analysis can be solved well
The above problem encountered in process.Substantive characteristics refers to the intrinsic feature of the target object unrelated with ambient enviroment, i.e. object
The style characteristic of reflection characteristic (the including the information such as color, texture) and object of body.Although for target object, the two
Inherent feature itself will not change with the variation of ambient enviroment, still, the observed image of collected target object
But it can be influenced by environment.Therefore, if the substantive characteristics of object can be analyzed directly from observed image, object has just been extracted
The information such as intrinsic shape, color, the texture of body eliminate influence of the environmental change to image, to realize to image more
Accurately understand, while also providing relatively reliable information base to further realize the image analysis of robust in complex environment
Plinth.
Current existing algorithm can be divided into three classes according to the mode of extraction target substantive characteristics:A kind of algorithm is implicit
Substantive characteristics parser, i.e., by algorithm for pattern recognition, to target object multi-modal, (target object is in different illumination conditions
And the presentation under different postures) learnt.During study, there is no the various presentations of special consideration for such algorithm
Between inner link, but directly to various observed results carry out pattern analysis, to attempt to obtain the target feature sky
Between in distribution function.The serious problems that such algorithm encounters are exactly the popularization problem of goal description function.Namely
It says, the distribution of training sample drastically influences the distribution function for finally learning to obtain.Exist if learning sample is target object
Image under single illumination conditions or posture, then the result of training study is difficult to be generalized to target object in different illumination conditions
Or on the image under new posture.Therefore, incomplete in sample, such algorithm is difficult to be generalized to target object to exist
Situation in various complex environments.Another kind of algorithm is explicit substantive characteristics parser.Such algorithm can be according to object
Idea's analysis of the body under different conditions inner link therein.Compared with first kind implicit algorithm, such algorithm is according to object
Image-forming principle and the priori of reflectance signature and shape are managed, analysis directly is made to the reflectance signature of object and shape.From
And image of target object in the state of new can directly be calculated according to these inherent features.Therefore, by explicit
The obtained result of parser it is more accurate, while also having more generalization.However, such algorithm generally use is based on knot
Estimation to substantive characteristics is realized in the constraints such as structure, texture, color, that is, follows the theoretical frame of Retinex by Modulation recognition problem
It is converted into energy-optimised problem, and completes to calculate under single scale and analyze.Thus, precision of analysis is largely
Dependent on the performance of optimization algorithm, simultaneously because it is impossible to ensure that the process that solves of the convexity of function to be optimized established often
Local minimum can be absorbed in and optimal solution can not be acquired, or require initialization step will be as possible close to optimal solution.These are all limited
The performance of such algorithm is made.Also a kind of algorithm is passed through using the essential image of neural network extraction based on deep learning
One convolutional neural networks of training, directly predict essential image from a width RGB image.But such existing algorithm network knot
Structure is fairly simple, and training set is by the artificial image of computer graphical software process quality, so the essential image of extraction is not
It is to be apparent from, when it being especially applied to natural image.
Invention content
In order to overcome existing implicitly and explicitly substantive characteristics parser ability in feature extraction insufficient, and it is based on depth
The problems such as neural network algorithm is mainly for artificial image is practised, the present invention provides a kind of with convolutional neural networks extraction essence figure
The method of picture.Build a double-current convolutional network with parallel construction from image to image first, then to the network into
Row training, optimizes network parameter, to extract the multilayer feature with environment invariance, directly reconstructs essential image (reflectogram
With illumination pattern).Using multiflow construction, on the one hand task can be made to detach, different features is extracted in different shuntings;Another party
Face, the two restrictive condition each other, can improve arithmetic accuracy.
A method of extracting essential image with convolutional neural networks, it is characterised in that steps are as follows:
Step 1:The double-current convolutional neural networks structural model with parallel construction is built, which is divided into a public affairs
There are branch, two proprietary branches.
Wherein, publicly-owned branch is made of 5 convolutional layers, and each convolutional layer is followed by a pond layer.The convolution kernel of convolutional layer
It is 3 × 3, each layer of one width characteristic image of output, it is 64 that the first convolutional layer, which exports characteristic image dimension, the output of the second convolutional layer
Characteristic image dimension is 128, and it is 256 that third convolutional layer, which exports characteristic image dimension, Volume Four lamination and the output of the 5th convolutional layer
Characteristic image dimension is 512, pond layer use size for 2 × 2 average pond.
Two proprietary branched structures are identical, separately include 3 warp laminations, and convolution kernel is 4 × 4, and a branch is used for
Light image is reconstructed, for reconstructing reflected image, the output dimension of all warp laminations is 256 for another branch.
The characteristic image of the third convolutional layer output of the publicly-owned branch is defeated with the second warp lamination of proprietary branch
Go out the input of the third warp lamination as proprietary branch;The publicly-owned branch Volume Four lamination output characteristic image with
Input of the output of first warp lamination of proprietary branch as the second warp lamination of proprietary branch.
Step 2:Training dataset is built, is cut by the middle part of every piece image of Jiang et al. BOLD data sets created
It is 1280 × 1280 image to take size, and respectively by five decile of interception image in row and column, then each in original data set
Width image obtains the image that 25 width sizes are 256 × 256, randomly selects 53720 groups of image construction test sets therein, residual graph
As composing training collection.
Step 3:The training set obtained using step 2 is trained the double-current convolutional neural networks that step 1 is built, first
Random initializtion is carried out to the weights of each layer of network, then using the training method for the error back propagation for having supervision, to network
It is trained, obtains trained network.Wherein, the basic learning rate of network is 10-13, using fixed learning rate strategy, network
Batch size be 5, loss function SoftmaxwithLoss, network convergence condition is the loss letter of front and back iteration twice
The difference of numerical value is in ± 5% range of its value.
Step 4:The test set obtained in step 2 is handled using trained network, the essence figure extracted
Picture, i.e. illumination pattern and reflectogram.
The present invention is also enterprising in essential image common data sets MIT Intrinsic Images dataset by this method
Go test, the results showed that, this method still has validity.
The beneficial effects of the invention are as follows:As a result of the technology road of the essential image zooming-out based on deep learning theory
Line can be directly from original image with the powerful ability in feature extraction of the neural network based on deep learning the Theory Construction
Isolate reflectogram and illumination pattern.Further it is proposed that double-current convolutional neural networks be a kind of complete from image to image
Convolutional network model is respectively used to generate illumination pattern and reflectogram including Liang Ge branches flow to;Also, the network structure will be compared with
Result of the high-rise convolution results after deconvolution operation is combined, and can enhance the details of characteristic pattern after deconvolution operation,
The reconstructed error for reducing illumination pattern and reflectogram to a certain extent, improves the ability of network characterization reconstruct.
Description of the drawings
Fig. 1 is a kind of method flow diagram extracting essential image with convolutional neural networks of the present invention
Fig. 2 is the double-current convolutional neural networks structure chart that the present invention is built
Fig. 3 is the data set parts of images example that the present invention is built
Specific implementation mode
Present invention will be further explained below with reference to the attached drawings and examples, and the present invention includes but are not limited to following implementations
Example.
The present invention provides a kind of methods for extracting essential image with convolutional neural networks, as shown in Figure 1, main process
It is as follows:
1, double-current convolutional neural networks structural model of the structure with parallel construction
The restructuring procedure of image is actually to assign different weights to the feature extracted from image, and by same type
Feature combines to complete to reconstruct the target of illumination pattern and reflectogram from original image.In other words, institute spy in need
Sign is all present in the same original image.Therefore characteristic extraction part can be shared, and two distinct types of essential image
Reconstruct then need separate completion.Thus, the network that the present invention is built is divided into two parts of public branch and proprietary branch.It is passing through
After crossing the convolution algorithm of publicly-owned branch, the characteristic pattern size of each layer output gradually reduces.In order to make input picture and output image
Same size is kept on space structure, is devised three warp laminations in two proprietary branches respectively, is made the sky of characteristic pattern
Between size be gradually restored to original size.It is inspired by residual error network structure, the present invention is had found during the experiment by publicly-owned point
Behind in branch two layers in proprietary branch behind two layers it is combined, network parameter can be made to obtain better effect of optimization.
Based on the above reason, the present invention constructs the double-current convolutional neural networks structure with parallel construction as shown in Figure 2.The network
Model is divided into a publicly-owned branch, two proprietary branches.
Wherein, publicly-owned branch is made of 5 convolutional layers, and each convolutional layer is followed by a pond layer.The convolution kernel of convolutional layer
It is 3 × 3, each layer of one width characteristic image of output, it is 64 that the first convolutional layer, which exports characteristic image dimension, the output of the second convolutional layer
Characteristic image dimension is 128, and it is 256 that third convolutional layer, which exports characteristic image dimension, Volume Four lamination and the output of the 5th convolutional layer
Characteristic image dimension is 512, pond layer use size for 2 × 2 average pond.Two proprietary branched structures are identical, respectively
Including 3 warp laminations, convolution kernel is 4 × 4, and for reconstructing light image, another branch is anti-for reconstructing for a branch
Image is penetrated, the output dimension of all warp laminations is 256.Also, the characteristic image of the third convolutional layer output of publicly-owned branch
The input of output with the second warp lamination of proprietary branch collectively as the third warp lamination of proprietary branch;Publicly-owned branch
The characteristic image of Volume Four lamination output is with the output of the first warp lamination of proprietary branch collectively as the second of proprietary branch
The input of warp lamination.
2, data set is built
Network structure proposed by the invention is more complicated, needs trained network parameter more.In order to make network play
Its optimal performance, BOLD data sets (Jiang X Y, Schofield AJ, the Wyatt J that the present invention is created in Jiang et al.
L.Correlation-Based Intrinsic Image Extraction from a Single Image[C]
.European Conference on Computer Vision,2010:One is constructed on the basis of 58-71) for studying
The data set of essential image zooming-out algorithm.The data set includes 268,600 groups of pictures, every group of picture include an original image,
One illumination pattern and a reflectogram.53,720 groups of composition test sets are therefrom extracted at random, for testing essential image zooming-out
Algorithm performance.Remaining 214,880 groups of composing training collection, for training deep learning neural network.BOLD databases include big
High-resolution image group is measured, they are all the objects shot under the lighting condition adjusted meticulously, include mainly various multiple
Miscellaneous decorative pattern, face and Outdoor Scene, Fig. 3 give data set parts of images example.Jiang et al. builds the purpose of the database
It is to provide a test platform for image processing algorithm.Specifically, including mainly essential image zooming-out algorithm, removing illumination algorithm
With light source algorithm for estimating etc..For this purpose, illumination condition figure and body surface figure, i.e. illumination pattern and reflectogram are they provided, and
All it is the standard rgb color space picture with linear luminance characteristic.The present invention is complicated from picture number, picture quality and scene
The various aspects such as degree consider, and final decision selection goes structure for this by intricate detail based on the picture group of reference object
The data set of research, original image have 1280 pixels in each dimension transverse direction, there is 1350 pixels on longitudinal direction, for
Data volume is excessively huge for common computer, and the problem of be easy to cause study, this is unfavorable for deep learning nerve very much
The training of network.There are one apparent features for the image category tool that the present invention chooses:Key message concentrates in the middle part of image.Cause
This, portion chooses one 1280 × 1280 feature frame to the present invention in the picture, intercepts original image, then in row and column respectively
By five decile of image.In this way, an original image is segmented into 25 256 × 256 smaller images.It is right in this way
Original image is cut, and the key message in original image is remained, and realizes data using maximization, while being also originally to grind
Study carefully and provides a variety of conveniences:The data volume of every group of picture is moderate, to computer performance without too high request;Relatively reasonable
Image size designs more convenient when convolutional neural networks;Image after shearing includes positive negative sample simultaneously, can be in certain journey
Over-fitting is avoided on degree.
3, network training
The present embodiment is under Caffe frames using constructed to train based on the training set in the data set that BOLD is created
Deconvolution neural network.It is compared with other frames, Caffe frames not only simple installation, but also all operating systems of support is also right
Python and Matlab has good interface to support.Since constructed network structure is relatively complicated, the data learnt are needed
To measure larger, it is more that network needs the number of iteration also to compare, and also to avoid e-learning is too fast from missing optimal solution, so
During being trained to network, the present invention determines basic learning rate being set as 10 by repetition test-13, learning rate plan
Slightly it is set as " fixed ", i.e., fixed learning rate.It is too fast also for network convergence is avoided in view of computer performance, network
Batch size are set to 5, loss function SoftmaxwithLoss.
Loss function is used for calculating increase of the difference between output result and true tag with iterations, network damage
Lose smaller and smaller, i.e., estimated result becomes closer to true tag.Loss function in single dimension can be write as:
Wherein, { (x1,y1),...,(xm,ym) indicate the good training data of m group echos, xIndicate that input, y indicate corresponding
Label, and y ∈ [0,255].1 { F } is indicator function, and functional value is 1 when F is genuine when be vacation, and functional value is 0, θ
Indicate the parameter of convolutional neural networks.In the training process, estimate that the error between image and true tag is backward propagated to
In neural network to optimize its parameter, so that error is gradually reduced.For RGB image, loss function is above-mentioned loss function
The sum of error in tri- dimensions of image R, G, B.
About after iteration 210,000 time, difference is floated in ± 5% range before and after loss function value, and network is gradually received
It holds back, i.e., tends to be optimal in Exist Network Structure lower network parameter, the ability of the essential image of network extraction tends to be best, is trained
Good network.Although although two proprietary branches look the same in structure, due to being supplied to their true tag not
Together, during network training, they learn obtained network parameter also can be different.Thus, extracting essence using network
When image, they also have corresponding difference to the operation of data so that different branches can extract different types of
Essential image.
4, essential image is extracted with trained network
The test set for including in the data set established in step 2 is handled using trained network, i.e., it will wherein
Including RGB pictures be converted into three-dimensional matrice, as the input of network, the essence extracted after the multilayer operation of network
Image, i.e. illumination pattern and reflectogram.The present invention is also by this method in essential image common data sets MIT Intrinsic
It is tested on Images dataset, the results showed that, this method still has validity.
Claims (1)
1. a kind of method for extracting essential image with convolutional neural networks, it is characterised in that steps are as follows:
Step 1:The double-current convolutional neural networks structural model with parallel construction is built, which is divided into one publicly-owned point
Branch, two proprietary branches;
Wherein, publicly-owned branch is made of 5 convolutional layers, and each convolutional layer is followed by a pond layer;The convolution kernel of convolutional layer is 3
× 3, each layer of one width characteristic image of output, it is 64 that the first convolutional layer, which exports characteristic image dimension, and the second convolutional layer exports feature
Image dimension is 128, and it is 256 that third convolutional layer, which exports characteristic image dimension, and Volume Four lamination and the 5th convolutional layer export feature
Image dimension is 512, pond layer use size for 2 × 2 average pond;
Two proprietary branched structures are identical, separately include 3 warp laminations, and convolution kernel is 4 × 4, and a branch is for reconstructing
Light image, for reconstructing reflected image, the output dimension of all warp laminations is 256 for another branch;
The characteristic image of third convolutional layer output of the publicly-owned branch is made with the output of the second warp lamination of proprietary branch
For the input of the third warp lamination of proprietary branch;The publicly-owned branch Volume Four lamination output characteristic image with it is proprietary
Input of the output of first warp lamination of branch as the second warp lamination of proprietary branch;
Step 2:Training dataset is built, is intercepted by the middle part of every piece image of Jiang et al. BOLD data sets created big
The small image for being 1280 × 1280, and respectively by five decile of interception image in row and column, then each width figure in original data set
The image for being 256 × 256 as obtaining 25 width sizes, randomly selects 53720 groups of image construction test sets therein, residual image structure
At training set;
Step 3:The training set obtained using step 2 is trained the double-current convolutional neural networks that step 1 is built, first to net
The weights of each layer of network carry out random initializtion, then using the training method for the error back propagation for having supervision, are carried out to network
Training, obtains trained network;Wherein, the basic learning rate of network is 10-13, using fixed learning rate strategy, network
Batch size are 5, loss function SoftmaxwithLoss, and network convergence condition is the loss function of front and back iteration twice
The difference of value is in ± 5% range of its value;
Step 4:The test set that step 2 obtains is handled using trained network, the essential image extracted, i.e. light
According to figure and reflectogram.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810407424.8A CN108764250B (en) | 2018-05-02 | 2018-05-02 | Method for extracting essential image by using convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810407424.8A CN108764250B (en) | 2018-05-02 | 2018-05-02 | Method for extracting essential image by using convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108764250A true CN108764250A (en) | 2018-11-06 |
CN108764250B CN108764250B (en) | 2021-09-17 |
Family
ID=64008978
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810407424.8A Active CN108764250B (en) | 2018-05-02 | 2018-05-02 | Method for extracting essential image by using convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108764250B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109658330A (en) * | 2018-12-10 | 2019-04-19 | 广州市久邦数码科技有限公司 | A kind of color development method of adjustment and device |
CN110659023A (en) * | 2019-09-11 | 2020-01-07 | 腾讯科技(深圳)有限公司 | Method for generating programming content and related device |
CN111179196A (en) * | 2019-12-28 | 2020-05-19 | 杭州电子科技大学 | Multi-resolution depth network image highlight removing method based on divide-and-conquer |
CN111325221A (en) * | 2020-02-25 | 2020-06-23 | 青岛海洋科学与技术国家实验室发展中心 | Image feature extraction method based on image depth information |
CN111489321A (en) * | 2020-03-09 | 2020-08-04 | 淮阴工学院 | Depth network image enhancement method and system based on derivative graph and Retinex |
CN113034353A (en) * | 2021-04-09 | 2021-06-25 | 西安建筑科技大学 | Essential image decomposition method and system based on cross convolution neural network |
US20210272236A1 (en) * | 2019-02-28 | 2021-09-02 | Tencent Technology (Shenzhen) Company Limited | Image enhancement method and apparatus, and storage medium |
CN114742922A (en) * | 2022-04-07 | 2022-07-12 | 苏州科技大学 | Self-adaptive image engine color optimization method, system and storage medium |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104756152A (en) * | 2012-10-26 | 2015-07-01 | Sk电信有限公司 | Image correction device for accelerating image correction and method for same |
CN106778583A (en) * | 2016-12-07 | 2017-05-31 | 北京理工大学 | Vehicle attribute recognition methods and device based on convolutional neural networks |
CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
CN107330451A (en) * | 2017-06-16 | 2017-11-07 | 西交利物浦大学 | Clothes attribute retrieval method based on depth convolutional neural networks |
CN107403197A (en) * | 2017-07-31 | 2017-11-28 | 武汉大学 | A kind of crack identification method based on deep learning |
CN107633272A (en) * | 2017-10-09 | 2018-01-26 | 东华大学 | A kind of DCNN textural defect recognition methods based on compressed sensing under small sample |
JP2018018422A (en) * | 2016-07-29 | 2018-02-01 | 株式会社デンソーアイティーラボラトリ | Prediction device, prediction method and prediction program |
-
2018
- 2018-05-02 CN CN201810407424.8A patent/CN108764250B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104756152A (en) * | 2012-10-26 | 2015-07-01 | Sk电信有限公司 | Image correction device for accelerating image correction and method for same |
JP2018018422A (en) * | 2016-07-29 | 2018-02-01 | 株式会社デンソーアイティーラボラトリ | Prediction device, prediction method and prediction program |
CN106778583A (en) * | 2016-12-07 | 2017-05-31 | 北京理工大学 | Vehicle attribute recognition methods and device based on convolutional neural networks |
CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
CN107330451A (en) * | 2017-06-16 | 2017-11-07 | 西交利物浦大学 | Clothes attribute retrieval method based on depth convolutional neural networks |
CN107403197A (en) * | 2017-07-31 | 2017-11-28 | 武汉大学 | A kind of crack identification method based on deep learning |
CN107633272A (en) * | 2017-10-09 | 2018-01-26 | 东华大学 | A kind of DCNN textural defect recognition methods based on compressed sensing under small sample |
Non-Patent Citations (3)
Title |
---|
GAO HUANG等: "Densely Connected Convolutional Networks", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
JIANG X Y等: "Correlation-Based Intrinsic Image Extraction from a Single Image", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
王珺等: "基于非下采样 Contourlet 变换和稀疏表示的红外与可见光图像融合方法", 《兵工学报》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109658330A (en) * | 2018-12-10 | 2019-04-19 | 广州市久邦数码科技有限公司 | A kind of color development method of adjustment and device |
CN109658330B (en) * | 2018-12-10 | 2023-12-26 | 广州市久邦数码科技有限公司 | Color development adjusting method and device |
US20210272236A1 (en) * | 2019-02-28 | 2021-09-02 | Tencent Technology (Shenzhen) Company Limited | Image enhancement method and apparatus, and storage medium |
US11790497B2 (en) * | 2019-02-28 | 2023-10-17 | Tencent Technology (Shenzhen) Company Limited | Image enhancement method and apparatus, and storage medium |
CN110659023A (en) * | 2019-09-11 | 2020-01-07 | 腾讯科技(深圳)有限公司 | Method for generating programming content and related device |
CN111179196A (en) * | 2019-12-28 | 2020-05-19 | 杭州电子科技大学 | Multi-resolution depth network image highlight removing method based on divide-and-conquer |
CN111179196B (en) * | 2019-12-28 | 2023-04-18 | 杭州电子科技大学 | Multi-resolution depth network image highlight removing method based on divide-and-conquer |
CN111325221A (en) * | 2020-02-25 | 2020-06-23 | 青岛海洋科学与技术国家实验室发展中心 | Image feature extraction method based on image depth information |
CN111325221B (en) * | 2020-02-25 | 2023-06-23 | 青岛海洋科技中心 | Image feature extraction method based on image depth information |
CN111489321A (en) * | 2020-03-09 | 2020-08-04 | 淮阴工学院 | Depth network image enhancement method and system based on derivative graph and Retinex |
CN113034353A (en) * | 2021-04-09 | 2021-06-25 | 西安建筑科技大学 | Essential image decomposition method and system based on cross convolution neural network |
CN114742922A (en) * | 2022-04-07 | 2022-07-12 | 苏州科技大学 | Self-adaptive image engine color optimization method, system and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN108764250B (en) | 2021-09-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108764250A (en) | A method of extracting essential image with convolutional neural networks | |
Liu et al. | Attribute-aware face aging with wavelet-based generative adversarial networks | |
CN112949565B (en) | Single-sample partially-shielded face recognition method and system based on attention mechanism | |
CN109815893B (en) | Color face image illumination domain normalization method based on cyclic generation countermeasure network | |
CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
CN111046964B (en) | Convolutional neural network-based human and vehicle infrared thermal image identification method | |
CN112699847A (en) | Face characteristic point detection method based on deep learning | |
Cao et al. | Ancient mural restoration based on a modified generative adversarial network | |
CN111489301B (en) | Image defogging method based on image depth information guide for migration learning | |
CN111445582A (en) | Single-image human face three-dimensional reconstruction method based on illumination prior | |
CN111861906B (en) | Pavement crack image virtual augmentation model establishment and image virtual augmentation method | |
CN102073995B (en) | Color constancy method based on texture pyramid and regularized local regression | |
CN110555908A (en) | three-dimensional reconstruction method based on indoor moving target background restoration | |
CN111402403B (en) | High-precision three-dimensional face reconstruction method | |
CN109635726B (en) | Landslide identification method based on combination of symmetric deep network and multi-scale pooling | |
CN113570658A (en) | Monocular video depth estimation method based on depth convolutional network | |
CN111507184B (en) | Human body posture detection method based on parallel cavity convolution and body structure constraint | |
CN110110782A (en) | Retinal fundus images optic disk localization method based on deep learning | |
CN113256494B (en) | Text image super-resolution method | |
CN114445268A (en) | Garment style migration method and system based on deep learning | |
CN117292117A (en) | Small target detection method based on attention mechanism | |
CN108388901B (en) | Collaborative significant target detection method based on space-semantic channel | |
CN111882516B (en) | Image quality evaluation method based on visual saliency and deep neural network | |
CN110727817B (en) | Three-dimensional model retrieval method based on t-CNN, terminal equipment and storage medium | |
CN113822825B (en) | Optical building target three-dimensional reconstruction method based on 3D-R2N2 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |